Fig. 4

PerSVade’s parameter optimization improves the SV calling accuracy on datasets with known real SVs. A To test perSVade’s performance on real SVs, we measured how the parameters optimized for several simulations in different species (see Fig. 3) work on three human samples (CHM, HG002, and NA12878) with defined sets of real SVs. Each row indicates one of these different “training” parameters optimized for each sample and simulation type. In addition, the first row refers to the default parameters. Each column represents a sample with defined real SVs to be “tested.” The heatmap shows the F-value of each parameter set on each tested real sample (hereafter referred to as “testing instance”). In addition, we divide the testing instances into different groups (“default,” “different spp,” “same spp,” and “same sample”), which are relevant to understand the B panel. The “different spp” group refers to instances where the training and testing species were different. The “~” (same spp) refers to instances where the training and testing samples were different, but from the same species. Finally, the “*” (same sample) refers to instances where the training and testing samples were the same. B We summarized the data shown in A to compare how similar types of training parameters performed on each testing sample (each represented by a different color). Each row corresponds to a different accuracy measure. Each point corresponds to a testing instance (matching one cell from the heatmap in A in the bottom “F-value” plots). The “default” and “same sample” reflect testing instances where the training parameters were either un-optimized or optimized specifically for each sample, respectively. The “different spp” group includes instances where the training parameters were from a different, non-human, species. The “same spp” group shows testing instances with both training parameters and tested simulations from different samples of the same species. In addition, each column represents testing instances where the training parameters were based on “random” or “known” simulations, respectively. Note that the different groups of “training parameters” are equivalent to those shown in A